You are on page 1of 6

2008 IEEE/RSJ International Conference on Intelligent Robots and Systems

Acropolis Convention Center

Nice, France, Sept, 22-26, 2008

System overview of bipedal robots Flame and TUlip:

tailor-made for Limit Cycle Walking
Daan Hobbelen, Tomas de Boer and Martijn Wisse

Abstract The concept of Limit Cycle Walking in bipedal Removing the need for dynamic balance results in a wider
robots removes the constraint of dynamic balance at every choice of walking motions and thus increased versatility.
instance during gait. We hypothesize that this is crucial for To demonstrate this potential of applying the Limit Cycle
the development of increasingly versatile and energy-effective
humanoid robots. It allows the application of a wide range of Walking paradigm in reality, we have built various physical
gaits and it allows a robot to utilize its natural dynamics in bipedal robots in the past and we are currently working on the
order to reduce energy use. This paper presents the design and construction of a new robot. This paper discusses the design
experimental results of our latest walking robot Flame and and experimental results of our latest walking robot, Flame,
the design of our next robot in line TUlip. The focus is on the and the design of the next robot in line, TUlip. Both robots
mechanical implementation of series elastic actuation, which is
ideal for Limit Cycle Walkers since it offers high controllability are depicted in Fig. 1. These robots are a first step towards a
without having the actuator dominating the system dynamics. versatile, energy-effective robot, as a combination of the typ-
Walking experiments show the potential of our robots, showing ically highly controllable ZMP based robots (showing poor
good walking performance, though using simple control. energy-effectiveness) and energy-effective Passive Dynamic
Walkers (showing limited versatility).
I. INTRODUCTION First we will discuss the design requirements that follow
Performing gait synthesis on bipedal robots is a useful way from the desire to maintain the advantageous principles
of increasing the understanding of the dynamic principles of Passive Dynamic Walking (e.g., natural dynamics, com-
of human walking. This understanding has two discretely pliance), while introducing sufficient actuation to make a
different types of application: robotics-oriented applications versatile system/research platform (Section II). Subsequently
and human-oriented applications. The robotics-oriented ap- Sections III, IV and V show a system overview of the two
plications include entertainment in the short-term [1] and robots, their actual mechanical realization and the control
home care in the long-term [2], [3]. The human-oriented ap- system architecture. As the first robot Flame has already
plications involve the improvement of rehabilitation devices, seen successfully operation, data from typical walking ex-
such as prosthetics and orthotics. periments is shown in Section VI. The paper ends with a
Our approach to bipedal gait synthesis uses the recently conclusion in Section VII.
introduced paradigm of Limit Cycle Walking [4]. This
paradigm is an extension on the concept of Passive Dynamic
Walking, as pioneered by McGeer [5]. Passive Dynamic Studying the potential benefits of Limit Cycle Walking on
Walkers show that it is possible to perform stable bipedal physical robots creates a set of special design requirements.
walking without any actuation or control. Limit Cycle Walk-
ing expands this concept to actuated bipeds. The essence of
Limit Cycle Walking is that it is possible to obtain stable
periodic walking without locally stabilizing the walking
motion at every instant during gait [4]. In this context, local
stability refers to stability in continuous-time for the direct
neighborhood of a state along a walkers motion trajectory.
Realizing that local stabilization is not necessary, eliminates
the need for dynamic balance (i.e., flat foot contact) as
introduced by Vukobratovic [6] and it allows the use of
compliant control. The possibility of using compliant control
makes that stiff control actions at energetically unfavorable
points along a motion trajectory can be avoided. Instead, a
Limit Cycle Walker can just stay close to its natural dynamics
at those locations (i.e., let go) and reject large disturbances by
control actions with low energy use, such as foot placement.

All authors are with the Delft Biorobotics Laboratory, Depart-

ment of BioMechanical Engineering, Faculty of Mechanical Engineer-
ing, Delft University of Technology, 2628 CD, Delft, The Netherlands Fig. 1. Walking robot Flame and design of the walking robot TUlip.

978-1-4244-2058-2/08/$25.00 2008 IEEE. 2486

As mentioned in the introduction, important advantageous
properties of Limit Cycle Walkers are that they move largely
according to their natural dynamics and that they have high
compliance/low control stiffness. A vital design requirement Flame TUlip
that results from this is that the actuators in the prototypes DOF Actuator Range of Actuation Range of
need to have a low output impedance, meaning that a type + reduction motion [rad] + reduction motion [rad]
Rx,ankle passive -0.05 to 0.05 passive -0.35 to 0.35
displacement imposed on the actuator by the walker will only Ry,ankle 90W, 1:223 -0.4 to 0.7 60W, 1:129 -0.6 to 0.6
result in a low actuator torque. In other words, the actuators Ry,knee 90W, 1:51 0 to 1.5 60W, 1:155 0 to 2.4
should leave the natural dynamics of the walker intact to be Rx,hip 90W, 1:416 0 to 0.2 60W, 1:111 - to 0.9
Ry,hip 90W, 1:223 -0.7 to 0.7 60W, 1:222 - to 0.9
able to properly apply the concept of Limit Cycle Walking. Rz,hip - - 60W, 1:86 -0.5 to
At the same time, in order to obtain high versatility and Ry,arm - - 20W, 1:411 2
the ability to handle large disturbances, the actuators should
supply a Limit Cycle Walker with sufficient controllability.
The actual requirements for versatility depend highly on
the research goals we have for the two prototypes. The
first prototype Flame is mainly designed to do straight A. Series Elastic Actuation
walking (no turns) at different speeds with the ability to To obtain the required combination of low actuator
start and stop. Its main research goal is the study of the impedance and high controllability, as discussed in Sec-
stabilizing effect of lateral foot placement. For the second tion II, the main choice of actuation for our robots is series
prototype TUlip the versatility goals have significantly been elastic actuation, as introduced by Pratt et al. [10]. This
expanded: it is designed to be able to do straight walking, actuator type is successfully implemented in various other
slow and hard turns and fall and stand up. These requirements robotic applications [11], [12].
follow from the desire to compete in the humanoid league The actuator system comprises a geared electric DC motor
of RoboCup [7]. For this purpose additional controllability which connects to the joint through an elastic element.
(additional active degrees of freedom), increased maximal By measuring the elongation of this element this actuation
torques and range of motion as well as increased design system allows the application of force/torque control as
robustness are required. shown by the schematic diagram in Fig. 3. The geared DC
Lastly, there are specific design requirements for the natu- motor is modeled as the motor input torque/current together
ral dynamics of the walkers, especially the natural dynamics with a reflected inertia and damping.
of the legs. From earlier studies we know that a naturally fast The use of electric DC motors generally allows high torque
swing-leg motion is desirable for both increasing walking control bandwidth with relatively little system overhead. In
speed, decreasing energy use [8] and increasing disturbance comparison to pneumatic or hydraulic actuators this makes
rejection [9]. The consequence of this for the walkers design the electric motor an attractive candidate for autonomous
is that we want the mass moment of inertia of the legs with robots. The application of a geared motor instead of direct
respect to the hip joint to be low. This can be achieved by drive is chosen because in case of direct drive the magnitude
minimizing the mass at the end of the legs. of the desired torques in our robots would require a heavy
motor with overly high power rating. However, the use of a
III. SYSTEM OVERVIEW geared electric DC motor typically results in a high actuator
In order for the robot to perform the required robotic tasks
as described in Section II, a combination of robotic joints arm
needs to be selected. For each robot, a minimal amount of
degrees of freedom is selected to minimize its complexity
and mass as much as possible. An overview of the degrees
of freedom of both robots is given in Fig. 2. Table I gives an
overview of the actuator type, reduction (by planetary gears)
and range of motion of each DOF (for the left side). knee

In Flame, lateral foot placement is implemented by means

of a compass like coupling mechanism (around Rx in the hip,
as shown in Fig. 2) that keeps the torso at half the inner leg Rz
angle. This is implemented for simplicity reasons to enable z y Ry
Flame TUlip
lateral foot placement controlling only one actuator. Since x
the required tasks of TUlip are more versatile compared to Rx

Flame, more degrees of freedom needed to be implemented

Fig. 2. Degrees of freedom of both Flame and TUlip. Flame and TUlip have
in the hip joint. Furthermore, two actuated arms are added a total of 9 respectively 14 degrees of freedom. The grey colored DOFs
that can be used during standing up. Both robots are 1.2 m are passive/non actuated DOFs. The sign of a DOF rotation is considered
tall and weigh 15 kg. with the robots upper body as the base.

0.4 0.4
geared electric motor With geared electric motor
Original (passive) dynamics

Hip pitch joint angle [rad]

bm,refl series elastic 0.2 0.2
joint 0 0

m Im,refl
-0.2 -0.2

With series elastic actuation

motor measured spring -0.4 -0.4
0 2 4 6 8 10 0 2 4 6 8 10
current/torque elongation/torque Time [s] Time [s]

controller Fig. 5. Measured hip pitch joint motion during passive swinging of a
leg (left), swinging with a geared electric motor directly connected (right,
Fig. 3. Schematic diagram of the concept of series elastic actuation. dotted) and with a series elastic actuator connected (right, dashed). Using
An electric motor drives a joint through an elastic/compliant element. By an elastic actuator leaves the original dynamics of the leg intact, in contrast
measuring the elongation of this element, the torque that the total actuator to the geared electric motor.
system delivers to the joint can be controlled.

tuator is clearly shown when comparing the system dynamics

impedance (i.e., high reflected inertia and damping/friction). before an actuator is connected to the dynamics with an ac-
This undesirable property is best solved by the application tuator as shown in Fig. 5. The left graph shows the measured
of torque control through a series elastic/compliant element dynamics of a freely swinging fully passive robot leg. The
as will be shown below. right graph shows how directly connecting a geared electric
Fig. 4 shows the modeled output impedance (actuator motor severely hampers these natural dynamics (dotted line).
torque over joint angular displacement) of one of the proto- In contrast, adding a series elastic actuator with (zero) torque
types actuators with and without the series elastic element controller leaves the original dynamics intact (dashed line).
plus torque controller present. The electric motor is a mod- Next to the low actuator impedance, series elastic actuation
eled as depicted in Fig. 3, for which the parameters have has another interesting advantage: shock/impact tolerance.
been fitted to measurements of the actual motor. The torque When impacts occur in the robots (and this does happen
controller is a proportional-derivative controller that has the frequently in walking), the series elasticity protects the motor
same settings as applied in the actual system. The high output and especially the gears in the gearbox. A disadvantage
impedance of a geared electric motor is shown by the dotted of the series elasticity is that it forms a limitation on the
line; a joint oscillation at 1 Hz with an amplitude of 1 rad position control bandwidth that can be obtained on the joints.
is resisted by a significant torque of up to 10 Nm, which is As indicated in the introduction, this is no problem when
close to the maximal torque the motor can deliver. In case applying Limit Cycle Walking as stability typically does not
series elastic actuation is applied, the actuator impedance is depend on the application of stiff position control.
reduced by at least a factor of 50 at all frequencies (dashed Finally, series elastic actuation enables measurement of the
and solid line). Using a more compliant element (instead of energy/torque on joint level, excluding the energy consumed
a typically stiff torque sensor) results in a lower actuator by the motor. This allows us to compare torque/energy of
impedance at high (uncontrolled) frequencies and a decrease our gait with other robotic or human gait, irrespective of the
in phase lead (i.e., less damping added by the actuator). motor type used.
The effect of low output impedance of a series elastic ac- B. Lower limb design
In general, substantial building space is required for series
elastic actuation, due to the required elastic element in the
drive train together with a sensor to measure the elongation
Magnitude (dB)

without series elastic actuation of the element. Therefore, series elastic actuation is only
50 (only geared electric motor) used in the joints where it is most beneficial, the joints
0 that have a highly dynamic (fast) motion during walking.
-50 with series elastic actuation increasing series These joints are the hip, knee and ankle pitch joints. The
practical implementation of series elastic actuation in those
180 joints is shown below, using the knee design as an example.
This example is followed by a description of foot and ankle
Phase (deg)

1) Exemplary knee design: Fig. 6 shows the implementa-
increasing series
tion of series elastic actuation in the knee joint of TUlip. We
0 compliance chose to implement series elastic actuation by using a tension
-1 0 1 2 3
10 10 10
Frequency (Hz)
10 10 spring in series with a steel cable. The use of cables offers
the possibility to place the actuator at a different position and
Fig. 4. Output impedance of one of our actuators with and without the orientation than the joint axis. The actuator is positioned as
presence of a series elastic element plus torque controller. high as possible in the upper leg to minimize the moment

idler general electronic hardware layout is described, followed by
the current control structure that is used to make Flame walk.
pulley A. Hardware layout
Both robots share a centralized control architecture where
DC all control is performed on a PC104 single board computer.
motor Control input data consists of motor and joint encoder
signals, inertial sensor data, foot contact switches and user
motor definable switches. Additionally, TUlip is fitted with two
encoder cameras and load sensors and accelerometers on the feet. All
data acquisition is performed using I/O cards (analog and/or
steel cable +
tension spring counters) which are stacked on the single board computer.
The control signals are sent to current control amplifiers that
encoder drive all motors.
The low power electronics and motor electronics are
powered separately by lithium polymer battery packs which
pulley can deliver a robot operating time of about 30 minutes. An
overview of the main control hardware components is given
in Table II.
B. Walking controller
Fig. 6. Schematic drawing of the left leg of TUlip together with a detailed
drawing of the implementation of series elastic actuation in the knee joint. Here we will explain the walking controller that is cur-
rently implemented on the robot Flame. With the controller
described here the walking results presented in Section VI
of inertia with respect to the hip joint. The orientation were obtained. The controller is run in software at 1 kHz
is axial along the leg, to enable placement inside the leg on the central PC104 processor. It has a cascade control
for protection. A 30000 counts per revolution incremental structure: the low-level (inner) control loop implements local
encoder is used on the load side for accurate position and torque control on each series elastically actuated joint and the
velocity estimation of the joint. By measuring the difference main controller supplies these local controllers with reference
in orientation of the encoder on the motor and load side, torques. The bandwidth of the torque controller is about 5-10
the extension of the spring can be determined, which is a Hz (bounded by system limitations, e.g. the resolution of the
measure for the torque exerted on the joint. velocity signal that is an input for the proportional-derivative
2) Foot and ankle design: In the ankle design, the only controller).
actuated joint is the ankle pitch joint. Again, the imple- The basis of the main controller is an event-based state
mentation of series elastic actuation using cables offers the machine that tracks which portion of the walking cycle the
possibility to place the ankle actuator away from the joint robot is in: double stance, left is leading leg, single stance,
itself. In this case the motor is placed in the upper body and left is stance leg, double stance, left is trailing leg or
is connected to the joint through a Bowden cable drive. This single stance, left is swing leg. The event that triggers the
construction follows the design requirement of keeping the transition from single stance to double stance is the detection
mass at the end of the leg low. It provides the possibility of
active plantarflexion of the ankle and dorsiflexion through a Bowden
return spring.
To achieve sideways stability, we believe that the effect of
return spring
lateral ankle actuation is limited due to the small foot width steel cable
resulting in a relatively low maximum torque. A minimal spring for + spring
lateral stability
design approach is chosen where a constant (but adjustable) (at both sides of foot)
ankle stiffness is implemented using only two tension springs
as shown in Fig. 7. joint joint
encoder Ry encoder Rx


This section will give a general overview of the control

system architecture that both robots share. Obviously, TUlip
will need more hardware (e.g. vision related) and more
sophisticated software than Flame in order to autonomously Fig. 7. Photo of the left foot of TUlip. Note that the foot electronics,
play a game of soccer. However, this is still under devel- comprising four load sensors and an accelerometer, are not shown on the
opment and not within the scope of this paper. Below, the photo since they are under development.

Flame TUlip
Computer 400 MHz Diamond Systems Athena board, 128 MB RAM 1 GHz Diamond Systems Poseidon board, 512 MB RAM
Operating system Linux + RTAI Linux + Xenomai
I/O cards Mesa Electronics 4I36 counter card (2x) Mesa Electronics 4I65 I/O card (2x)
Diamond-MM-16-AT 16-bit analog I/O card
Inertial sensor XSens MTi XSens MTi
Joint encoder Scancon 2RMHF 30000 counts/rev Scancon 2RMHF 30000 counts/rev
Motor encoder Agilent 2000 counts/rev Agilent 2000 counts/rev
Vision - Custom, under development
Foot sensor Contact switch Tekscan Flexiforce load cell + accelerometer
Motor Maxon RE 35 Maxon RE 30, RE 25 in arms
Amplifier AMC Z12A8 6 A RMS Elmo Whistle 3.5 A RMS
Electr. battery Kokam 3 cell 6 Ah LiPo Kokam 3 cell 6 Ah LiPo
Motor battery Kokam 8 cell 26.4 Ah LiPo Kokam 8 cell 26.4 Ah LiPo
User input Switches Matrix Orbital LK204-25 LCD

of swing heel strike by the foot sensor. The transition from VI. WALKING EXPERIMENTS
double stance to single stance is triggered by the end of
Preliminary results of walking experiments with Flame are
push-off (the ankle push-off torque gets below a threshold
shown as a first indication that the Limit Cycle Walking
concept can generate stable gait using simple control and
For the ankle pitch joint, both in single and double stance, has potential to be used for increased versatile and energy-
the main controller generates reference torques that emulate effective humanoid robots.
a spring stiffness. In the leading leg in double stance and The internal joint motions and joint torques measured in
stance leg in early single stance, the spring stiffness is low ten steps of a typical (unperturbed) walking experiment at a
to induce soft landing of the foot with limited bouncing speed of approximately 0.45 m/s are shown in Fig. 8. The
after impact [13]. In late stance and in the trailing leg during solid lines give the information for the stance/leading leg, the
double stance the stiffness is high. Moreover, in the trailing dotted lines for the swing/trailing leg. The grey area indicates
leg in double stance, the equilibrium angle of the emulated the duration of the double stance phase and the white area the
spring is discretely changed to provide propulsion by push- single stance phase. To complete the orientation information
off. In the swing leg the ankle pitch actuation ensures that for the robot, Fig. 9 gives the absolute orientation of Flames
the toes are temporarily lifted to prevent toe stubbing. upper body with respect to gravity.
For the knee pitch joint, the main controller has two Next to these unperturbed walking experiments, we also
functions: in the stance leg it generates a reference torque that performed walking experiments in which Flame was delib-
pushes the knee in its hyperextension stop to prevent knee erately perturbed. An example of a typical perturbation that
collapse, in the swing leg it ensures bending and stretching of Flame can handle is an 8 mm stepdown in the floor. This
the knee to create ground clearance. The latter action requires experiment is shown in an accompanying short video.
little torque as the bending and stretching motions are close
to the natural dynamic knee motion. VII. CONCLUSION
The controller for the hip pitch joint also has two main The walking robot Flame and the design of the latest robot
functions: the stance leg hip pitch joint torque keeps the prototype TUlip were presented, showing the evolution of
upper body upright through proportional-derivative feedback our Limit Cycle Walkers towards more versatile humanoid
from the inertial sensor on the upper body, the swing leg robots. The combination of Limit Cycle Walking control
hip pitch joint torque ensures that the swing leg is brought with the implementation of series elastic actuation has great
forward to a desired fixed inter-leg angle. As the natural potential for future bipedal robots, as preliminary, successful
dynamics of the leg are not fast enough this takes active walking experiments indicate.
acceleration and deceleration, mainly obtained by a feed-
forward hip pitch torque pattern and regulated by low gain VIII. ACKNOWLEDGMENTS
feedback. This research is funded by the Dutch Technology Foun-
The hip roll joint has no series elastic actuation and thus dation STW, applied science division of NWO and the
has a direct position control task. Towards the end of the Technology Program of the Ministry of Economic Affairs,
single stance phase the hip roll joint regulates proper lateral the Commission of the European Union, within Framework
swing foot placement in order to obtain stability of the Programme 6, RTD programme IST, under contract no. FP6-
robots lateral motion [14]. The swing foot placement is 2005-IST-61-045301-STP and the 3TU.Centre for Intelligent
based on a linear strategy that only uses information of Mechatronic Systems. The authors would like to thank Garth
the lateral position and velocity of the robots hip. This Zeglin and all employees and students of the 3TU.Federation
information is acquired through the robots inertial sensor. universities that have contributed to this research.

Joint angles [rad] Joint torques [Nm]
0.5 keeping upper body upright
decelerate swing leg
Hip pitch

0 knee impact

-0.5 double single stance accelerate swing leg
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
1 knee bending 20
Knee pitch

knee extending
10 low torques in swing

-10 overstretching torque in stance

0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
0.5 foot landing push-off loading ankle spring
Ankle pitch


-0.5 push-off toe up (prevent stubbing) 'soft' landing low torques in swing
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
Hip roll angle [rad]

lateral foot placement

prevent foot collision
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Time [s]

Fig. 8. Measured internal joint angles and torques in ten typical steps of the robot Flame. The solid lines give the information for the stance/leading leg,
the dotted lines for the swing/trailing leg.
Body pitch [rad]

Body roll [rad]

0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
Time [s] Time [s]

Fig. 9. Measured upper body pitch and roll angle in ten typical steps of the robot Flame.

R EFERENCES plest walking model. ASME Journal of Biomechanical Engineering,

124(2):113120, 2002.
[1] T. Ishida. Development of a small biped entertainment robot QRIO. In [9] M. Wisse, A. L. Schwab, R. Q. v. d. Linde, and F. C. T. v. d. Helm.
Proceedings of International Symposium on Micro-Nanomechatronics How to keep from falling forward; elementary swing leg action for
and Human Science, 2004. passive dynamic walkers. IEEE Transactions on Robotics, 21(3):393
401, 2005.
[2] K. Hirai. Humanoid Robot and its Applications. In Proceedings
[10] G. A. Pratt and M. M. Williamson. Series Elastic Actuators. Pro-
of First International Workshop on Humanoid and Human Friendly
ceedings of IEEE International Conference on Intelligent Robots and
Robotics 1998, International Advanced Robotics Program, 1998.
Systems, 1995.
[3] K. Yokoi, N. Kawauchi, N. Sawasaki, T. Nakajima, S. Nakamura,
[11] J. Pratt, C. -M. Chew, A. Torres, P. Dilworth and G. Pratt. Virtual
K. Sawada, I. Takeuchi, K. Nakashima, Y. Yanagihara, K. Yokoyama,
Model Control: An Intuitive Approach for Bipedal Locomotion. The
T. Isozumi, Y. Fukase, K. Kaneko, and H. Inoue. Humanoid Robots International Journal of Robotic Research, 20(2):129-143, 2001.
applications in HRP. In Proceedings of International Conference on
[12] J. F. Veneman, R. Ekkelenkamp, R. Kruidhof, F. C. T van der Helm
Humanoid Robots, 2003.
and H. A. van der Kooij. A Series Elastic- and Bowden-Cable-Based
[4] D. G. E. Hobbelen and M. Wisse. Limit Cycle Walking. In M. Hackel, Actuation System for Use as Torque Actuator in Exoskeleton-Type
editor, Humanoid Robots, Human-like Machines, book chapter 14. I- Robots The International Journal of Robotics Research 25:261-281
Tech Education and Publishing, 2007. 2006 .
[5] T. McGeer. Passive Dynamic Walking. The International Journal of [13] D. G. E. Hobbelen and M. Wisse. Ankle Actuation for Limit Cycle
Robotics Research, 9(2):6282, 1990. Walkers. The International Journal of Robotics Research, 27(6): 709-
[6] M. Vukobratovic and B. Borovac. Zero-Moment Point - thirty years of 735, 2008.
its life. International Journal on Humanoid Robotics, 1(1):157174, [14] D. G. E. Hobbelen and M. Wisse. Active lateral foot placement for 3D
2004. stabilization of a Limit Cycle Walker prototype. International Journal
[7] RoboCup website, of Humanoid Robotics (in review), 2008.
[8] A. D. Kuo. Energetics of actively powered locomotion using the sim-